On-line Classifier for UAV Pilot Identification

  • Sultan AlAmeri

Student thesis: Master's Thesis


Unmanned Arial Vehicles (UAVs) have been a major global interest for the past few years. This interest has been ranged between commercial and military applications. Those applications initiated long term investments by multiple governments internationally to focus on the UAV research field from different aspects. One of these research fields covers the enhancement of UAV wireless link in terms of signal security and encryption. After all, UAVs rely on a wireless communication link associated occasionally with a ground station for telemetry and flight path controlling purposes. Thus, the UAV is vulnerable to hijacking by either intercepting its wireless communication link or by physically taking over its ground station. On either case, the hijacking can be stopped by deactivating the UAV fully at certain conditions. This deactivation is determined by the successive type of commands received by the UAV based on the pilot behavior analysis. The pilot behavior can be analyzed and detected by using different machine learning classifiers. Multiple classifying algorithms can be used for such goal such as support vector machine, random forest, and quadratic discriminant [1]. Those algorithms provide different range of classifying accuracy depending on the nature of the dataset being processed. In case of a UAV controlled by a remote controller, four main control signals are processed to define a pilot behavior. Those are pitch, roll, yaw, and thrust signals. A preliminary study, evaluating how accurate each algorithm classifying a pilot behavior based on those four features, shows that random forest performs relatively better than the other ones [1][2]. Random forest is a machine learning classifier based on the voting of several decision trees built by the bootstrapping process [3]. Bootstrapping is the process of randomly sampling two thirds of the entire dataset to train and construct the threshold of each node of a decision tree. Each decision tree depth can be varied ranging from shallow to deep depending on the complexity of the dataset and the features that are being trained on. Also, the number of constructed decision trees plays a significant role for defining the accuracy of pilot behavior identification results. These two key points become critical when developing an online classifier for a real-time system where processing time and memory size are involved.
Date of AwardNov 2017
Original languageAmerican English
SupervisorAbdulhadi Shoufan (Supervisor)


  • Machine Learning
  • Random Forest
  • Online Classifier
  • Real-time
  • UAV.

Cite this